Introducing TRC Computer Race Ratings
Total Page：16
File Type：pdf, Size：1020Kb
Introducing TRC Computer Race Ratings By James Willoughby Computers are very, very good at handicapping racehorses. If a computer is taught everything a human knows about the process, a human has no shot whatsoever of producing a set of ratings that finds more winners than a computer. None. In the old world, people used to cling to the concept that human intuition and intelligence could defeat computers, even at tasks that had a finite state of possibilities like chess and Go. But it just turned out that the computers had not been programmed adequately. Humans are good at handicapping too. We know that because the ratings that computers produce aren’t all that different from humans. But computers are faster, more accurate and never get tired. Until computer vision is refined, humans are still needed to collect the data about races germane to the task of handicapping. For example, a human might tell a computer which horses were hampered – and by how much – and which horses were eased at the end of a race. To really sharpen computer handicapping, a human might encode variables for the ease a horse was travelling or how hard-ridden it was. Humans can also be useful for providing computers with so-called ‘priors’ – estimates about things the computer has not otherwise been taught and which it benefits from as a starting point. The problem with computers is that they are just dumb robots. They carry out tasks according to their programmer’s instructions and no more. Although some impressive tasks that computers can carry out in the real world are labelled as ‘intelligent’, they aren’t intelligence as we possess it. Yet. But computers can mimic learning by iterative processes and by chaining vast networks of simple processing units together capable of understanding massively complex patterns in data. Teaching the TRC computer to handicap This article is not going to descend into mathematical details, which are abstruse to most followers of the sport. While there is no avoiding them if you are building a handicap model, it is relatively easy to understand the concept of computer handicapping explained in words. Maths is, however, a concise, beautiful language all of its own and talking maths to a computer bridges the communication gap, so here is how we started the process of teaching the TRC computer to handicap horses: There were 156,906 performances in the TRC database, covering Group and Graded races from January 5, 2011, to February 9, 2021. First, the computer needs to know the obvious: the distance a horse is beaten is a function of the quality of a race and the racing merit of the horse. And it’s the latter we are trying to rate. We also need to tell it that horses don’t always perform to the same level. Far from it. Their performance – measured in distance beaten where seconds or lengths are the units – actually follows a pattern called the gamma distribution. They come in at intervals behind the winner, with the greatest concentration coming just after the winner and the gaps tending to get wider as the back markers trail home. It is interesting to ponder why this might be the case. A ten-runner horse race, for instance, isn’t decided by ten individual time trials. If it were, we could describe the finishing times by a normal distribution, such as the one that describes adult male heights, for instance. But a horserace is different because of tactics. In races where the winner goes for home in good time, lesser-talented rivals have to go hard in pursuit before it is really good for them. So, as you go down the field, each horse tends formerly prominent to run less efficiently if it were trying to win the race. It’s a different matter with closers who run through beaten horses in their own time. While the gamma distribution is well described in statistics, it can be a bit of a nightmare to deal with. The normal distribution is better behaved, and it is desirable for our purposes to take the distances behind the winner and somehow ‘normalise them’, that is, transform them mathematically so that when we stack up all the figures in the database, the pattern looks like adult male heights. That’s what the first line of the model above says to the computer. The fancy ψ is just the Greek letter Psi used here both to denote a ‘normalising function’. It says: “Take the distances (or time lags) and manipulate them so that they follow a normal distribution. The tilde ‘~’ is read “distributed as’ and the N means the normal distribution. But the model also says do something else. Before you normalise the distances, flip them about because our convention is that higher-rated horses are better and thus get beaten shorter distances. And turn lengths (or seconds) into rating points using a formula: So, distance beaten is now a score which for the latest Breeders’ Cup Mile looked like this: Horse DTW PTW SCORE Order Of Australia IRE 0.00 0 17 Circus Maximus IRE 0.37 1 16 Lope Y Fernandez IRE 1.49 5 12 Ivar BRZ 2.99 9 8 Uni GBR 3.36 10 7 Halladay USA 4.11 11 6 Kameko USA 4.86 13 4 Factor This USA 5.98 15 2 Siskin USA 7.47 17 0 Raging Bull FRA 8.59 19 -2 Digital Age IRE 9.71 20 -3 Casa Creed USA 10.46 21 -4 March To The Arch USA 13.82 24 -7 Safe Voyage IRE 22.41 29 -12 mean=> 14 What we are doing is plugging the distance to the winner (DTW) into the above formula to produce rating points to the winner (PTW). Can you see the benefits of the formula? It expands the distances close behind the winner and shrinks those further away. Because the further behind a horse finishes, the less significant is a length or two. The 45 in the formula caps the maximum ratings points a horse can be defeated, while the ratio between 45 and 12 of just less than 4 determines how valuable the first length is (over a mile, remember.) The final column takes the mean of PTW, which is 14, and subtracts each individual PTW from it. After this, the score is normalised (the process is beyond the scope of this article) and fixed in the database. Leaning the effect of every course, distance and going on beaten lengths So far, so good. Let’s look at the way we taught the computer to handicap again: The picture to the right of the equations is a typical normal distribution. Its shape is described by two parameters, the mean μ (Greek letter Mu) which is the location, or middle, of the range of possible values, which itself is described by the standard deviation σ (Greek letter Sigma). The vertical dashed lines show the value on the horizontal axis corresponding to one standard deviation less (to the left of the mean) and more (to the right of the mean) than μ. The computer first takes a stab at the ratings of all races and horses. It tries to find the set of ratings for all races and horses that has the highest likelihood, given the data. In other words, the ratings of all horses and races that are most consistent with the results. Adjusting for conditions For every performance, the difference between the horse rating and the race quality is the distance the computer expects the horse to be beaten. But, when the actual distance doesn’t match, the computer has to figure out why. We tell it to expect that the variability from the distance it expects is a function of the course, the distance and the country-specific going (‘Good’ in France is very different going from ‘Good’ in Japan, for instance.) We won’t go over the importance of this consideration yet again. You can read about why we think it leads to horses in the Far East, in particular, being underrated in this article. The computer thus has an allowance for every set of circumstances a horse can encounter. Here is a portion of the chart that we can extract from the model: These are not the exact allowances that we use in all circumstances, merely the mean of those allowances. If a race that is run under conditions that normally stretch finishers out has already resulted in wide margins between the runners, we won’t extend the distances still further. But the allowances in the chart above give a decent means of comparison between conditions. The larger margins we allow for races in Japan and Hong Kong are only partly responsible for horses from those countries receiving standout ratings under our system; a larger part is their outstanding record internationally. So, now the computer has more than 150,000 equations to solve simultaneously. Do you remember doing two of these at school? Because there are far fewer unknowns than equations - 46,000 horses and 16,000 races - we say the system of equations is ‘overdetermined’, which means there are many possible values for every horse in every country. The relative strength of racing in each country is partly determined by the performance of challengers from other countries: Leaning not just the most likely ratings for all horses and races but also their range of values and the probabilities of these values requires software and some serious processing. All the while, the computer is learning how to handicap by tuning the values it comes up with according to how well they predict future races.